fix(reputation): never cadence-downgrade a trusted submitter, and never claim a review was skipped when it ran

src/queue/processors.ts

@@ -8284,9 +8284,18 @@ export function maybeAddReputationSkipHold(
     confirmedContributor: boolean;
     skipAiReview?: boolean | undefined;
     reputationSkipped: boolean;
+    /** True when this pass is a forced re-run, which bypasses the reputation skip -- see body. */
+    forceAiReview?: boolean | undefined;
   },
 ): boolean {
   if (!args.reputationSkipped || !shouldRequirePublicAiReviewForAdvisory(env, args)) return false;
+  // The hold must describe what HAPPENED, not what the heuristic wanted: aiReviewAllAuthors and a forced
+  // re-run (the panel checkbox / `@loopover review`) both bypass the reputation skip in
+  // shouldStartAiReviewForAdvisory, so on those passes the AI review genuinely runs. Pushing the "was
+  // skipped" finding anyway produced a published surface that asserted the review was skipped DIRECTLY
+  // ABOVE the review's own summary (observed live on JSONbored/loopover#9764) -- and, worse, held the gate
+  // for a review that succeeded.
+  if (args.settings.aiReviewAllAuthors || args.forceAiReview) return false;
   args.advisory.findings.push({
     code: "ai_review_inconclusive",
     severity: "warning",
@@ -11042,6 +11051,7 @@ async function maybePublishPrPublicSurface(
       confirmedContributor,
       skipAiReview: webhook.skipAiReview,
       reputationSkipped: preComputedReputationSkip === true,
+      forceAiReview: webhook.forceAiReview,
     });
     // #9035: a caught reviewer-manipulation attempt is evidence of intent, not a code-quality observation, and
     // must not buy an automated decision. Same fail-closed shape as the two holds above.

src/review/reputation-wire.ts

@@ -127,6 +127,15 @@ export async function shouldSkipAiForReputation(
   // upgrade-only AMS bridge inside getEffectiveSubmitterReputation, then cadence (#4514) independently.
   const stats = await getEffectiveSubmitterReputation(env, { repoFullName: args.project, submitter: args.submitter });
   if (shouldDowngradeToDeterministic(stats)) return true;
+  // A TRUSTED submitter is never cadence-downgraded. The machine-paced leg is a spend control against
+  // UNPROVEN accounts flooding paid review; "trusted" is this same module's own definition of proven (5+
+  // recent merged successes at a low fail rate, recency-windowed -- see signalFromCounts). Without this, a
+  // maintainer running a legitimate high-velocity campaign (observed live: the repo OWNER, 50 submissions in
+  // 24h at a 7.2-minute median gap, 145 recent merges) had EVERY PR downgraded to deterministic-only and
+  // held for manual review -- the gate taxing exactly the contributor with the strongest earned record. An
+  // abuser cannot reach "trusted" cheaply: it requires real merged PRs on this repo inside the window, and
+  // any prompt-injection row hard-locks the signal to "low" regardless.
+  if (stats.signal === "trusted") return false;
   const cadence = await getSubmitterCadence(env, args.project, args.submitter ?? undefined);
   return isMachinePacedCadence(cadence);
 }

test/unit/auto-review-wiring.test.ts

@@ -459,6 +459,44 @@ describe("review.auto_review wiring (#1954)", () => {
     ).toBe(false);
   });
 
+  it("REGRESSION: no 'was skipped' hold on a FORCED re-run — the bypass means the review actually ran", () => {
+    // Observed live on JSONbored/loopover#9764: the published panel asserted "Required AI review was skipped
+    // by a submitter-reputation downgrade" DIRECTLY ABOVE the very AI summary that forced re-run produced,
+    // and held the gate for a review that succeeded. forceAiReview bypasses the reputation skip in
+    // shouldStartAiReviewForAdvisory, so the hold must describe what happened, not what the heuristic wanted.
+    const blockingEnv = { AI_SUMMARIES_ENABLED: "true", AI_PUBLIC_COMMENTS_ENABLED: "true", AI: {} } as Env;
+    const settings = { gatePack: "oss-anti-slop", aiReviewMode: "block", aiReviewAllAuthors: false } as never;
+    const advisory = { headSha: "sha", findings: [] as unknown[] };
+    const added = maybeAddReputationSkipHold(blockingEnv, {
+      settings,
+      advisory: advisory as never,
+      repoFullName: "acme/widgets",
+      author: "the-maintainer",
+      confirmedContributor: false,
+      reputationSkipped: true,
+      forceAiReview: true,
+    });
+    expect(added).toBe(false);
+    expect(advisory.findings).toEqual([]);
+  });
+
+  it("REGRESSION: no hold under aiReviewAllAuthors either — that setting bypasses the skip the same way", () => {
+    const blockingEnv = { AI_SUMMARIES_ENABLED: "true", AI_PUBLIC_COMMENTS_ENABLED: "true", AI: {} } as Env;
+    const settings = { gatePack: "oss-anti-slop", aiReviewMode: "block", aiReviewAllAuthors: true } as never;
+    const advisory = { headSha: "sha", findings: [] as unknown[] };
+    expect(
+      maybeAddReputationSkipHold(blockingEnv, {
+        settings,
+        advisory: advisory as never,
+        repoFullName: "acme/widgets",
+        author: "anyone",
+        confirmedContributor: false,
+        reputationSkipped: true,
+      }),
+    ).toBe(false);
+    expect(advisory.findings).toEqual([]);
+  });
+
   it("#9015: a reputation skip HOLDS where blocking AI review is required — suspicion must never buy less scrutiny", () => {
     const blockingEnv = { AI_SUMMARIES_ENABLED: "true", AI_PUBLIC_COMMENTS_ENABLED: "true", AI: {} } as Env;
     const settings = { gatePack: "oss-anti-slop", aiReviewMode: "block", aiReviewAllAuthors: false } as never;

test/unit/reputation-wiring.test.ts

@@ -339,6 +339,36 @@ describe("shouldSkipAiForReputation (helper)", () => {
       expect(await shouldSkipAiForReputation(env, { project: "acme/widgets", submitter: "speedster" })).toBe(true);
     });
 
+    it("REGRESSION: a TRUSTED submitter is never cadence-downgraded, no matter how fast they pace", async () => {
+      // Observed live on the ORB: the repo OWNER, mid remediation campaign (50 PRs in 24h, 7.2-minute median
+      // gap, 145 recent merges), had EVERY PR downgraded to deterministic-only and held for manual review.
+      // The machine-paced leg is a spend control against UNPROVEN accounts; "trusted" is this module's own
+      // definition of proven (5+ recent merged successes, low fail rate), so it must short-circuit cadence.
+      const env = createTestEnv({ LOOPOVER_REVIEW_REPUTATION: "true" });
+      // 6 merged outcomes in the window -> signal "trusted" (trustedMinSuccess=5, failRate 0).
+      for (let i = 0; i < 6; i++) {
+        await seedReviewTarget(env, { project: "acme/widgets", repo: "acme/widgets", number: 700 + i, installationId: 1, submitter: "the-maintainer", status: "merged", reasonCode: "success" });
+      }
+      // ...and a blatantly machine-paced burst: 8 PRs at a 4-minute gap, all inside the cadence window.
+      const t0 = Date.now() - 60 * 60_000;
+      for (let i = 0; i < 8; i++) {
+        await seedCadencePullRequest(env, { number: 800 + i, submitter: "the-maintainer", createdAt: new Date(t0 + i * 4 * 60_000).toISOString() });
+      }
+      expect(await shouldSkipAiForReputation(env, { project: "acme/widgets", submitter: "the-maintainer" })).toBe(false);
+    });
+
+    it("INVARIANT: the trusted exemption cannot mask genuine abuse — a prompt-injection row still downgrades", async () => {
+      // "trusted" is unreachable with a prompt-injection row in the window: signalFromCounts hard-locks the
+      // signal to "low" before any success counting. This pins that ordering, so the exemption above can
+      // never become a loophole where an abuser banks merges and then injects at machine pace.
+      const env = createTestEnv({ LOOPOVER_REVIEW_REPUTATION: "true" });
+      for (let i = 0; i < 6; i++) {
+        await seedReviewTarget(env, { project: "acme/widgets", repo: "acme/widgets", number: 900 + i, installationId: 1, submitter: "sleeper", status: "merged", reasonCode: "success" });
+      }
+      await seedReviewTarget(env, { project: "acme/widgets", repo: "acme/widgets", number: 990, installationId: 1, submitter: "sleeper", status: "closed", reasonCode: "source_prompt_injection" });
+      expect(await shouldSkipAiForReputation(env, { project: "acme/widgets", submitter: "sleeper" })).toBe(true);
+    });
+
     it("FLAG-ON: false for the same number of submissions spread naturally over hours (comfortably human pace)", async () => {
       const env = createTestEnv({ LOOPOVER_REVIEW_REPUTATION: "true" });
       const t0 = Date.now() - 20 * 60 * 60_000;